Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Generalized Venn and Venn-Abers Calibration with Applications in Conformal Prediction

T0 review · 4 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Venn and Venn-Abers calibration can be defined for any loss function, and any point predictor that is perfectly calibrated in-sample can be turned into a prediction set that always contains a marginally calibrated point in finite samples.

desk verdict Elegant generalization of Venn calibration, but the quantile-loss conformal guarantee rests on a derivative condition that isotonic regression does not satisfy. read the letter →

arxiv 2502.05676 v3 pith:W5R2HMXI submitted 2025-02-08 stat.ML cs.LGstat.ME

Lars van der Laan , Ahmed Alaa This is my paper · ORCID
classification stat.MLcs.LGstat.ME MSC 62G0562G0862G20
keywords VenncalibrationVenn-Abersconformalpredictionmulticalibrationisotonicregressiondistribution-freeset-valuedquantile
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a unified framework for Venn and Venn-Abers calibration that works with any loss function, not just binary classification. The core result is that any point calibrator which is perfectly calibrated on the calibration data can be transformed into a set-valued predictor: for a new input, the method outputs a set of predictions guaranteed to contain one point that is perfectly calibrated marginally over the calibration data in finite samples. As the sample grows, the set shrinks and converges to a single conditionally calibrated prediction, so the method behaves like a point calibrator while quantifying epistemic uncertainty in small samples. The same principle is extended to multicalibration across subpopulations, and for quantile loss it recovers group-conditional and multicalibrated conformal prediction as special cases, yielding prediction intervals with quantile-conditional coverage.

What carries the argument

The Venn prediction set is the central object. For a new context $x$, the algorithm runs a point calibration algorithm $A_\ell$ on the calibration data augmented once for each possible imputed outcome $y$, and collects the predictions $f_n^{(x,y)}(x)$ for that context into one set. The identity that carries the argument is that the oracle member—the one obtained by imputing the true outcome $Y_{n+1}$—lies in this set by construction; exchangeability then converts the point calibrator's in-sample calibration on the augmented data into a marginal calibration guarantee for that member. For the quantile-loss application, the derivative of the quantile loss, $\partial\ell_\alpha(f(x), S(z)) = (1-\alpha) - \mathbf{1}\{S(x) \geq f(x)\}$, is what connects calibration of the conformity-score quantile to coverage of the resulting interval.

What would settle it

Run Algorithm 2 on a continuous outcome with a finite imputation grid that excludes the true test outcome; if the interval then fails to achieve exactly $1-\alpha$ coverage conditional on the calibrated quantile (or if the true point is not in the set), the finite-sample guarantee stated in Theorem 3.1 does not apply to the implemented discrete version.

Watch

Extended reading notes

Core claim

The central discovery is Theorem 3.1: under exchangeability (C1), finite second moment (C2), and perfect in-sample calibration of the point calibrator (C3), the Venn prediction set $f_{n,X_{n+1}}(X_{n+1})$ always contains the oracle member $f^*_{n+1}(X_{n+1}) = f_n^{(X_{n+1},Y_{n+1})}(X_{n+1})$, which is marginally perfectly $\ell$-calibrated: $\mathbb{E}[\mathbb{E}[\partial\ell(f^*_{n+1}(X_{n+1}), Z_{n+1}) \mid f^*_{n+1}(X_{n+1})]^2] = 0$. In words, imputing every possible outcome $y$, recalibrating on the augmented dataset, and collecting all predictions for the new context yields a set that always contains a prediction whose calibration error is zero marginally—an exact finite-sample guarantee where point calibrators can only offer asymptotic ones. The same construction with the quantile loss turns the set into a conformal interval with conditional coverage, and the multicalibration version gives finite-sample calibration across subpopulations.

Load-bearing premise

The proof requires that the point calibrator is perfectly calibrated in-sample on every augmented dataset—including the one containing the new input and its true outcome—and for generic losses this condition is asserted rather than proven, with the squared-error isotonic case being the one fully worked out.

Editorial extensions

If this is right

  • Any point calibrator that achieves perfect in-sample calibration—histogram binning, isotonic regression, or regression trees—automatically yields a set-valued predictor with a finite-sample marginal calibration guarantee.
  • As the calibration sample grows, the Venn prediction set shrinks to a single point, so the set-valued predictor asymptotically matches the conditional calibration of the underlying point calibrator.
  • Applying the framework to quantile loss produces conformal prediction intervals with quantile-conditional coverage: conditional on the calibrated quantile, coverage is exactly $1-\alpha$ in finite samples.
  • Venn multicalibration provides a finite-sample guarantee that the prediction set contains a prediction that is calibrated across every subpopulation in a given function class, and it recovers multicalibrated conformal prediction as a special case.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the discretization of a continuous outcome space uses a grid that does not contain the true test outcome, the oracle member is not in the finite set, so the exact finite-sample guarantee in Theorem 3.1 fails; the method then becomes approximate, and the size of the approximation error is not quantified in the paper.
  • The same 'impute and recalibrate' construction could be applied to other data perturbations—bootstrap resampling or leave-one-out—suggesting a general recipe for converting any stable calibrator into a set-valued uncertainty quantification tool, although the paper does not pursue this.
  • The paper's computational note that the prediction set's range can often be found by iterating over only the extreme outcomes $\{y_{\min}, y_{\max}\}$ when the mapping $y \mapsto f_n^{(x,y)}(x)$ is monotone suggests that in monotone settings the full set is an interval, which would make the method practical for continuous outcomes; the paper does not state a formal monotonicity theorem.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a unified framework for Venn and Venn-Abers calibration for generic loss functions. Under exchangeability, it shows that the prediction set obtained by recalibrating on every imputed test outcome contains an oracle point prediction that is marginally perfectly calibrated in finite samples (Theorem 3.1), and that the set collapses asymptotically to a conditionally calibrated point (Theorems 3.2 and 3.4). It extends the construction to multicalibration over finite-dimensional function classes (Theorem 3.5), and applies the framework to conformal prediction with quantile loss, claiming quantile-conditional coverage guarantees (Theorems 4.1 and 4.2) and connecting to the multicalibrated conformal prediction framework of Gibbs et al. (2023). Numerical experiments on several real datasets are reported for quantile Venn-Abers CP and for regression Venn multicalibration.

Significance. If the central result held in the advertised generality, it would give a clean finite-sample marginal-calibration guarantee for set-valued predictors under exchangeability, unifying several calibration and conformal prediction constructions; the paper also ships code and states falsifiable claims, which is a strength. The core exchangeability step in Theorem 3.1 is mathematically clean conditional on the in-sample calibration condition C3. However, the practical significance is weakened by three issues: the guarantee is existential because the calibrated member of the set is not identified; the continuous-outcome implementation discretizes Y so the oracle prediction is generally not contained in the computed set; and the quantile-loss applications rely on an in-sample differentiability condition that isotonic quantile regression does not satisfy. The asymptotic conditional-calibration results are also not fully proved in the appendix.

major comments (4)
  1. [Section 4.1, Theorem 4.1 and Algorithm 2] The quantile-loss application is not supported by the framework as written. The proof of Theorem 4.1 passes from C10 (empirical optimality over transformations) to the derivative condition C3 used in Theorem 3.1, but for the non-smooth quantile loss this passage requires differentiability of the empirical objective at the isotonic solution. For a constant block of an isotonic quantile fit, the minimizer is an order statistic (or, in the integer-alpha case, an interval endpoint), so the sum of the paper's derivatives (1-alpha) - 1{S_i >= q} is not zero for generic alpha and block size; selecting an interior point of the argmin interval does not fix this unless alpha n is exactly an integer. The no-tie assumption S_i != f*_{n+1}(X_i) is violated at the order-statistic observation when the standard PAVA solution is used, and if instead an interior point is chosen, the derivative-free optimality of the nonsmooth objective is lost. Appendix B establishes perfect in-sample calibration only for squared-error isotonic regression, not for quantile loss. Consequently, Theorem 4.1's conclusion P(S_{n+1} <= f*_{n+1}(X_{n+1}) | f*_{n+1}(X_{n+1})) = 1 - alpha is not established, and the same gap invalidates the claimed quantile-conditional coverage in Theorem 4.2 and in the Section 5.1 experiments; the exchangeability argument alone supports only the usual split-conformal marginal lower bound.
  2. [Section 3.2, Computational considerations] Theorem 3.1 requires the Venn set to contain the imputation with the true test outcome Y_{n+1}, which requires iterating over all y in the outcome space. The manuscript immediately states that for non-discrete outcomes Algorithm 2 is approximated by iterating over a finite subset of Y with linear interpolation, and the experiments use 200 equal-frequency bins for Y. Unless the true Y_{n+1} falls exactly on the grid, the oracle prediction f^{(X_{n+1},Y_{n+1})}_n(X_{n+1}) is not a member of the computed set, so the finite-sample marginal-calibration guarantee does not hold for the implemented method. The claim of 'negligible approximation error' is not quantified and does not follow from any stated condition. The paper should either restrict the finite-sample theorem to discrete or gridded outcomes with a guarantee that the grid contains the true label, or provide a quantitative bound on the coverage loss due to discretization.
  3. [Appendix C.1, Theorems 3.2 and 3.4] The asymptotic conditional-calibration results are not fully proved in this manuscript. The proof of Theorem 3.2 contains an explicit placeholder 'epsilon-covering number (?)' and uses notation bdelta_n that is never defined; the proof of Theorem 3.4 is delegated to prior self-cited work (van der Laan et al. 2023, 2024a) and refers to 'Lemma 3', which does not appear anywhere in the present paper. Since Theorem 3.4 is used to support the claim that Venn-Abers sets collapse asymptotically to a conditionally calibrated point and to justify the CCE comparisons in Section 5.1, the derivation needs to be completed in the appendix or the result should be stated as conditional on a fully specified theorem from the cited papers.
  4. [Section 5.1, Table 1] The experimental evaluation does not test the quantity appearing in Theorem 4.1. The CCE in Section 5.1 is defined conditional on f*_n, the isotonic calibration of f trained on the calibration set C_n, whereas Theorem 4.1's guarantee is stated for the oracle prediction f*_{n+1} = f^{(X_{n+1},Y_{n+1})}_n trained on the augmented data including the test point. The interval bC_n(X_{n+1}) uses the threshold f^{(X_{n+1},y)}_n for each y, but the reported conditional calibration error conditions on a different predictor. The empirical CCE and width results should be either re-derived for the oracle member of the Venn set or clearly presented as a heuristic evaluation of the point-calibrated baseline.
minor comments (4)
  1. [Theorem 3.3] The statement lists only C1 and C2, but it is asserted to follow from Theorem 3.1, which also requires C3. The statement should include the appropriate in-sample calibration condition, otherwise it is technically false in the stated generality.
  2. [Section 3.3, Algorithm 3] The shortcut of iterating over the extreme points {y_min, y_max} is justified only 'under monotonicity of y -> f^{(x,y)}_n', but the paper does not state conditions under which Algorithm 3's linear-offset calibrator is monotone in the imputed outcome. For general linear regressions this monotonicity is not automatic and should be checked or assumed explicitly.
  3. [Section 4.1, C10] The condition C10 is formulated as an empirical risk equality over transformations, while Theorem 3.1 uses the pointwise derivative condition C3. These are not equivalent for nonsmooth losses, and the proof of Theorem 4.1 should make the link explicit rather than simply saying the result follows from Theorem 3.1.
  4. [Throughout] There are several typos and OCR artifacts, including 'T ovk' in the abstract, 'Averaged averaged' in Section 5.2, 'Shermen and Morrison' in the references, and '1D functions' in the proof of Theorem 3.2; these should be cleaned up in revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Venn set contains the oracle by transparent construction, and the marginal-calibration claim is a non-trivial exchangeability argument, not a renaming of assumptions.

full rationale

The central claim (Theorem 3.1) is not circular. Algorithm 1 defines fn,x(x) := {f_n^{(x,y)}(x) : y in Y}, and Theorem 3.1 identifies f*_{n+1}(X_{n+1}) = f_n^{(X_{n+1},Y_{n+1})}(X_{n+1}) as a member of that set; the paper states this membership 'by construction'. The substantive part of the theorem is that this oracle member is marginally perfectly ℓ-calibrated, i.e., E[E[∂ℓ(f*,Z_{n+1})|f*]^2] = 0. That conclusion is derived from C3 (in-sample calibration of f* on the augmented n+1 sample) together with exchangeability, via the standard symmetrization step: the in-sample derivative sums are equal in expectation to the test-point derivative because f* is permutation-invariant. This is a genuine probabilistic transfer, not a restatement of C3. The quantile applications (Theorems 4.1-4.2) explicitly assume C10 / C9 rather than deriving them from Algorithm 2; whether isotonic quantile regression satisfies these in-sample conditions is a correctness/assumption gap, not circularity. The proof of Theorem 3.4 cites the authors' prior isotonic-calibration results as proof templates; those are external published theorems and not equivalent to the present claim, so they do not make the derivation circular. Section 5.2's 'oracle Venn-calibrated prediction' uses the true test label by design and is labeled as such; this is a limitation of the experimental illustration, not a hidden fitted-input prediction. Finally, the paper's own computational notes acknowledge that continuous Y is handled on a finite grid, which weakens the finite-sample guarantee in practice; again this is a practical approximation, not a circular step.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The framework rests on exchangeability and on the in-sample calibration condition C3, which is an assumption about the point calibrator rather than a parameter fitted to data. The user-specified hyperparameters (bin counts, discretization grid, spline basis) are listed as free parameters.

free parameters (3)
  • number of histogram bins K = not specified in theory; 200 used in experiments
    The point calibrators (histogram binning, isotonic regression) require choosing bin granularity; the theoretical rates depend on k(n). Not fitted to data, but a user-chosen hyperparameter.
  • discretization grid size for Y = 200 equal-frequency bins (Sections 3.2, 5.1)
    Algorithm 2 is approximated by discretizing continuous Y into 200 bins; the approximation error is not quantified.
  • spline basis for multicalibration function class G = cubic splines with five knot points (Section 5.2)
    The function class G in Algorithm 3 is user-specified; the experiments choose an additive cubic spline basis.
assumptions (6)
  • domain assumption C1: Exchangeability of (X_i,Y_i) for i=1..n+1
    Invoked in Theorem 3.1 and used throughout to equate the test point with calibration points.
  • domain assumption C3: Perfect in-sample calibration of the point calibrator on the augmented dataset
    The central premise of Theorem 3.1; must hold exactly for the chosen A_l. Shown for squared-error isotonic regression in Appendix B, assumed for general losses.
  • domain assumption C2: Finite second moment of the loss derivative
    Technical condition for the conditional expectation argument in Theorem 3.1.
  • domain assumption C4-C7: i.i.d., boundedness, Lipschitz derivative, finite number of bins
    Conditions for the asymptotic conditional calibration rate in Theorem 3.2.
  • ad hoc to paper C8: Finite total variation of the conditional gradient
    Imposed specifically to obtain the n^{-2/3} rate for isotonic calibration in Theorem 3.4; not verified in experiments.
  • standard math Standard empirical process bounds (uniform entropy integrals, peeling)
    Used in proofs of Theorems 3.2 and 3.4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalized Venn and Venn-Abers Calibration with Applications in Conformal Prediction." pith.science (2026). https://pith.science/paper/W5R2HMXI

@misc{pith2026250205676,
  author       = {Pith},
  title        = {Pith review of: Generalized Venn and Venn-Abers Calibration with Applications in Conformal Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W5R2HMXI}},
  note         = {Machine review of arXiv:2502.05676}
}
read the original abstract

Ensuring model calibration is critical for reliable prediction, yet popular distribution-free methods such as histogram binning and isotonic regression offer only asymptotic guarantees. We introduce a unified framework for Venn and Venn-Abers calibration that extends Vovk's approach beyond binary classification to a broad class of prediction problems defined by generic loss functions. Our method transforms any perfectly in-sample calibrated predictor into a set-valued predictor that, in finite samples, outputs at least one marginally calibrated point prediction. These set predictions shrink asymptotically and converge to a single conditionally calibrated prediction, capturing epistemic uncertainty. We further propose Venn multicalibration, a new approach for achieving finite-sample calibration across subpopulations. For quantile loss, our framework recovers group-conditional and multicalibrated conformal prediction as special cases and yields novel prediction intervals with quantile-conditional coverage.

Figures

Figures reproduced from arXiv: 2502.05676 by the authors.

Figure 1
Figure 1. Prediction bands capturing epistemic uncertainty in the calibration of point predictions generated by our proposed methods with a squared error loss. In safety-critical sectors such as healthcare, it is crucial to ensure that model-driven decisions are reliable under mini￾mal assumptions (Mandinach et al., 2006; Veale et al., 2018; Vazquez and Facelli, 2022; Gohar and Cheng, 2023). Point calibrators, which map a sin… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Fitted Occupancy-Ratio Evaluation without Bellman Completeness

    stat.ML 2026-07 conditional novelty 7.0 of 10

    FORE estimates discounted occupancy ratios by iterating KL-projected adjoint Bellman updates, achieving convergence under ratio realizability alone without Bellman completeness.

Reference graph

Works this paper leans on

21 extracted references · 12 canonical work pages · cited by 1 Pith paper

  1. [5]

    Multicalibration: Calibration for the (computationally-identifiable) masses

    Ursula H´ebert-Johnson, Michael Kim, Omer Reingold, and Guy Rothblum. Multicalibration: Calibration for the (computationally-identifiable) masses. In International Conference on Machine Learning , pages 1939–1948. PMLR,

    work page 1939

  2. [10]

    Calibration of probabilities: The state of the art

    Sarah Lichtenstein, Baruch Fischhoff, and Lawrence D Phillips. Calibration of probabilities: The state of the art. In Decision Making and Change in Human Affairs: Proceedings of the Fifth Research Conference on Subjec- tive Probability, Utility, and Decision Making, Darmstadt, 1–4 September, 1975, pages 275–324. Springer,

    work page 1975

  3. [12]

    Lars van der Laan, Ziming Lin, Marco Carone, and Alex Luedtke

    PMLR. Lars van der Laan, Ziming Lin, Marco Carone, and Alex Luedtke. Stabilized inverse probability weighting via isotonic calibration. arXiv preprint arXiv:2411.06342, 2024a. Lars van der Laan, Alex Luedtke, and Marco Carone. Au- tomatic doubly robust inference for linear functionals via calibrated debiased machine learning. arXiv preprint arXiv:2411.027...

    arXiv 2011

  4. [13]

    Fairness and accountability design needs for algorithmic support in high-stakes public sector decision-making

    Michael Veale, Max Van Kleek, and Reuben Binns. Fairness and accountability design needs for algorithmic support in high-stakes public sector decision-making. In Pro- ceedings of the 2018 chi conference on human factors in computing systems, pages 1–14,

    work page 2018

  5. [15]

    Orthogonal causal calibration

    Justin Whitehouse, Christopher Jung, Vasilis Syrgkanis, Bryan Wilder, and Zhiwei Steven Wu. Orthogonal causal calibration. arXiv preprint arXiv:2406.01933,

    arXiv

  6. [18]

    is a data-adaptive histogram binning method that learns the bins using isotonic regression, a nonparametric method traditionally used for estimating monotone functions (Barlow and Brunk, 1972; Groeneboom and Lopuhaa, 1993). Specifically, the bins are selected by minimizing an empirical MSE criterion under the constraint that the calibrated predictor is a ...

    work page 1993

  7. [19]

    By assumption, γf (θ(x,y) n , X)∂ℓ(θ(x,y) n (f (x)), y) is uniformly bounded, such that 1 n nX i=1 γf (θ(x,y) n , Xi)∂ℓ(θ(x,y) n (f (Xi), Yi)) = O(n−1). Adding and subtracting, we have that Pnγf (θ(x,y) n , ·)∂ℓ(θ(x,y) n (f (·), ·)) = O(n−1) P γf (θ(x,y) n , ·)∂ℓ(θ(x,y) n (f (·), ·)) + (Pn − P )γf (θ(x,y) n , ·)∂ℓ(θ(x,y) n (f (·), ·)) = O(n−1) P {γf (θ(x,...

    work page 2011

  8. [20]

    Proof of Theorem 3.4

    Thus, bδ2 n = Op(δ2 n) and the result follows. Proof of Theorem 3.4 . This proof follows from a generalization of the proofs of Theorem 1 for treatment effect calibration and propensity score calibration in van der Laan et al. (2023) and van der Laan et al. (2024a). Recall that f (x,y) n = θ(x,y) n ◦ f. Under C8, up to a change of notation, the proof of L...

    work page 2023

Show all 21 references
  1. [21]

    Taking the expectation of both sides, which we can do by C2, and leveraging C1, we find that 0 = 1 n + 1 n+1X i=1 E g(Xi) Yi − f ∗ n+1(Xi) = 1 n + 1 n+1X i=1 E g(Xn+1) Yn+1 − f ∗ n+1(Xn+1) = E g(Xn+1) Yn+1 − f ∗ n+1(Xn+1) . C.3. Proofs for conformal prediction Proof. Proof of ...

    2023

  2. [1996]

    Con- formal prediction with conditional guarantees

    Isaac Gibbs, John J Cherian, and Emmanuel J Cand`es. Con- formal prediction with conditional guarantees. arXiv preprint arXiv:2305.12616,

    arXiv

  3. [2002]

    11 Generalized Venn and Venn-Abers Calibration A. Code Availability Python code implementing Venn-Abers and Venn multicalibration methods for both squared error and quantile losses is available in the VennCalibration package at the following GitHub repository: https://github.c...

    2002

  4. [2005]

    The scope of multicalibra- tion: Characterizing multicalibration via property elicita- tion

    10 Generalized Venn and Venn-Abers Calibration Georgy Noarov and Aaron Roth. The scope of multicalibra- tion: Characterizing multicalibration via property elicita- tion. arXiv preprint arXiv:2302.08507,

    arXiv

  5. [2008]

    Batch multivalid conformal prediction

    Christopher Jung, Georgy Noarov, Ramya Ramalingam, and Aaron Roth. Batch multivalid conformal prediction. arXiv preprint arXiv:2209.15145,

    arXiv

  6. [2011]

    Efficient venn predictors using random forests

    Ulf Johansson, Tuve L¨ofstr¨om, Henrik Linusson, and Henrik Bostr¨om. Efficient venn predictors using random forests. Machine Learning, 108:535–550, 2019a. Ulf Johansson, Tuwe L¨ofstr¨om, and Henrik Bostr¨om. Cali- brating probability estimation trees using venn-abers pre- dic...

    2019 arXiv

  7. [2012]

    Venn-abers predictors.arXiv preprint arXiv:1211.0025,

    Vladimir V ovk and Ivan Petej. Venn-abers predictors.arXiv preprint arXiv:1211.0025,

    arXiv

  8. [2016]

    ISBN 978-1-4503- 4232-2

    ACM. ISBN 978-1-4503- 4232-2. doi: 10.1145/2939672.2939785. URL http: //doi.acm.org/10.1145/2939672.2939785. Joel W Cohen, Steven B Cohen, and Jessica S Banthin. The medical expenditure panel survey: a national information resource to support healthcare cost research and infor...

  9. [2017]

    Happymap: A generalized multi-calibration method

    9 Generalized Venn and Venn-Abers Calibration Zhun Deng, Cynthia Dwork, and Linjun Zhang. Happymap: A generalized multi-calibration method. arXiv preprint arXiv:2303.04379,

    arXiv

  10. [2018]

    Smooth isotonic regression: a new method to calibrate predictive models

    Xiaoqian Jiang, Melanie Osl, Jihoon Kim, and Lucila Ohno- Machado. Smooth isotonic regression: a new method to calibrate predictive models. AMIA Summits on Transla- tional Science Proceedings, 2011:16,

    2011

  11. [2022]

    Mul- tiaccuracy: Black-box post-processing for fairness in classification

    Michael P Kim, Amirata Ghorbani, and James Zou. Mul- tiaccuracy: Black-box post-processing for fairness in classification. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society , pages 247–254,

    2019

  12. [2023]

    A survey on intersectional fairness in machine learning: Notions, mitigation, and challenges

    Usman Gohar and Lu Cheng. A survey on intersectional fairness in machine learning: Notions, mitigation, and challenges. arXiv preprint arXiv:2305.06969,

    arXiv

  13. [2024]

    Forster-warmuth counterfactual re- gression: A unified learning approach

    Yachong Yang, Arun Kumar Kuchibhotla, and Eric Tch- etgen Tchetgen. Forster-warmuth counterfactual re- gression: A unified learning approach. arXiv preprint arXiv:2307.16798,

    arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.